On May 15, AMD officially unveiled Trinity, the company’s successor to its Llano line of "Fusion" processors that combine CPU and GPU into a single part. The company announced five models of the new processor: three for "mainstream" notebook computers, desktops, and "all-in-one" systems; and two targeted at what AMD calls the "ultrathin" notebook market.
With a 17-watt version available for "ultrathin" notebooks, AMD can offer up an alternative to Intel’s trademarked Ultrabook. "In theory, AMD has something that can work there, but we'll have to see how it plays out in the real world," said David Kanter, Manager and Editor of Real World
Technologies, in an interview with Ars.
And there’s one other wrinkle to deal with: Intel’s pending delivery of its low-voltage Ivy Bridge processor this summer. "The big question is, from a power and performance standpoint, how does the 17-watt Trinity play out relative to Ivy Bridge?" Kanter said. "We haven’t seen the Ivy Bridge parts for Ultrabook yet."
The layout of the Trinity "Accellerated Processor Unit", or APU.
Piling on the Piledriver
Trinity is a step forward from an architectural standpoint—albeit not a long one. The CPU side of Trinity is based on an architecture code-named Piledriver, the successor to AMD’s (somewhat disappointing) Bulldozer architecture. Trinity processors have two or four Piledriver cores, alongside the new Radeon 6000 GPU (with up to 384 Radeon cores) that takes up the other half of the die. (Oddly, Radeon is branding the Radeon 6000 in Trinity as a Radeon 7000.)
Piledriver inherits a number of things from Bulldozer. First, there’s shared L2 cache for each pair of cores, as opposed to the private L2 cache used in Llano. And the L1 cache for each core is "write-through" just as in Bulldozer—as opposed to Llano’s write-back cache. That makes for slower writes to cache, but somewhat faster reads.
There are some evolutionary changes from Bulldozer in Piledriver. AMD executives said that there have been modifications made to the branch prediction used in Piledriver, but didn’t get into specifics. There’s more bandwidth to access L2 cache in Piledriver, and larger L1 cache Translation Lookaside Buffers (TLBs) for each core, which make it less likely that their memory manager will waste processing cycles doing a "page walk" searching for memory addresses.
Other improvements that come with Piledriver are additions to the instruction set that was missing from Bulldozer. When Bulldozer was originally being developed, Intel made changes to how it was implementing the Fused Multiply-Add (FMA) extension to the x86 instruction set, which speeds up floating-point calculations. So Bulldozer shipped with the older FMA4 version of the implementation (the 4 standing for "four operands"). The version chosen by Intel to be implemented in the company’s Haswell processors, which will be introduced in 2013—FMA3— has been added to Piledriver, for compatibility.
One notable improvement is a resonant clock mesh technology, licensed from Cyclos Semiconductor that allows the processor to boost its clock speed in a more energy-efficient way. The resonant clock system kicks in when the processor runs in modes faster than 2.9 GHz, and will allow future AMD processors to reach clock speeds over 4 GHz. As a result, Piledriver’s architecture is more energy efficient than Llano at high clock speeds.
But Kanter said that at least part of the better performance and power profile of Trinity comes from the overall system design, and the processor’s ability to shift power dynamically between the CPU and GPU based on demand. Llano’s GPU lacked any real power management of its own.
Good news, bad news
All of that adds up to a significant improvement over its Llano predecessors. It also makes Trinity more competitive with a wide swath of Intel’s Sandy Bridge CPUs, though just how competitive depends heavily on how you measure them.
AMD claims Trinity delivers up to a 29 percent increase in CPU performance for its current top-end Trinity processor (the A10-4600M, a 4-core processor with a maximum clock-rate of 3.2 GHz and 4 megabytes of L2 cache) over the Llano A8-3500M, and over 56% better GPU performance. AMD also claims that Trinity gets twice the CPU performance per Watt of Llano.
Then there are the comparisons to Intel's current notebook processor. Trinity’s programmable GPU is certainly more powerful than that of Intel’s Sandy Bridge. And overall—if you count the GPU—AMD claims Trinity’s raw computing capacity in gigaflops is six times that of Intel’s mid-range i5-2520M (though that processor is a dual core, it runs the same number of threads as the Trinity chip). AMD also claimed that the power management capabilities of Trinity gave systems based on it better battery life than an equivalent Sandy Bridge-based notebook.
Right now, AMD’s Trinity has a big advantage over Intel’s Ivy Bridge in the low-power ultra-whatever range: Trinity is shipping. (So, apparently, is AMD's Brazos 2.0 low-power chip, the company's competitor to Atom, though AMD spoke about Brazos 2.0 briefly and in the future tense at the Trinity event). According to AMD executives, over a million Trinity parts have already been shipped. Intel won’t start shipping ultrabook-friendly Ivy Bridge processors until this summer, though Ivy Bridge is shipping for desktops and less-than-ultra notebooks.
But the benchmark news needs to be taken with a very substantial grain of salt. ExtremeTech’s Joel Hruska called all these claims into doubt because of conflicting benchmarks within AMD’s own results, and suggested that Trinity’s performance was about on par with a Llano processor of the same clock speed. And regardless of how you spin the numbers, Trinity still lags behind Intel’s processors in terms of CPU performance—the kind of computing power that most applications depend on (despite AMD’s efforts to get people to build applications that leverage its GPU).
Given the price range that AMD is aiming for, some of the benchmarks may not have a lot of meaning since they’re usually run on the highest-end chips. "Reviews rarely capture what the low-end and mid-range of the market looks like," Kanter said, "but that is where most chips are sold." Yes, core for core, Trinity is not going to nuke Sandy Bridge or Ivy Bridge on performance—but considering that it's priced so aggressively lower than Intel's sort-of-comparable processors (AMD is shooting for ultrabook-ish systems for under $500), that might not be much of an issue.
And in the realm of the 17-watt "ultrathin", there are other issues beyond the processor that can have as much of an impact on the computer’s perceived performance. But just how much help AMD’s graphics prowess will be for the company on low power systems isn’t clear—right now, the main "killer app" for graphics processors is gaming, and low-power ultra-whatevers are not exactly the gaming platform of choice. (Then again, if you were looking for a gaming ultra-thin notebook, Trinity might be the CPU for you.)
"AMD’s advantage in graphics depends on the power envelope," Kanter said. "The larger the envelope, the larger their advantage." So Trinity may end up having more hearty adoption on the desktop, where people appreciate its graphics chops more.
